typical AI startup sprint, fixed scope
Service AI startup development · UK seed-to-Series-A
AI startup development for the seed founder who needs the eval pass before the term sheet.
You’re pre-Series-A with your wedge, and your VC’s technical partner wants eval pass rates, idempotent tools, drift monitoring, and the AI off-switch in code. That’s the build we ship in twelve to sixteen weeks with two senior engineers paired daily.
24hreply, from a senior
200+projects shipped since 2019
Senioronly, on the spine
Your VC asks two questions. Does the model work, and what happens when you turn it off.
That’s the call you’re walking into at the next board meeting. We build the AI startup product so both answers already exist in code, written down, defendable in plain English. Three things carry the whole conversation.
Eval, in writing.
Held-eval suite plus a regression gate in CI. The pass rate gets published in the term-sheet appendix, so your investor reads a number, not a promise.
Idempotent by design.
Every tool call is replay-safe. No double-charge, no double-email on a retry. The failure mode your CFO worries about is architecturally impossible.
The off-switch.
Your product still earns its keep if the AI quota runs out tomorrow. We design the fallback before we wire the model, not after the outage.
Nadia raised seed. Then her VC asked three questions she couldn’t answer.
She raised seed in 2024 for an AI-first revenue ops platform. Two engineers, 6 weeks to first paying customer. Then her VC’s technical partner sat in on the next board call. Here’s how it went, and where it landed.
What’s the eval pass rate, and where’s the audit log?
What happens to customers if OpenAI goes down for two hours. Where is the per-inference audit log. What is the held-eval pass rate. Nadia’s team had none of these. The board pulled the Series A bridge.
Eval suite, model registry, failover, audit log.
We rebuilt in fourteen weeks. 240 golden Q&A pairs gated in CI. Model registry on MLflow. Idempotent tool calls. Bedrock plus Azure failover. An audit log per inference, defendable in plain English.
The bridge closed. The Series A closed inside the quarter.
The next board call closed the bridge, and the Series A closed inside the quarter at a 3.4x post-launch valuation lift. For the founder who decided the AI is the product floor and wants the eval pass rate in writing.
Six things. Every one of them in the term-sheet appendix.
You don’t get a demo and a deck. You get the six artefacts a Series A technical partner asks for, built into the product from week one, each one a line your investor can verify on a call.
01
Eval suite + regression gate
Held-eval golden Q&A pairs. A regression breaks the build. Pass rate published per release.Golden Q&A setCI regression gateBraintrust / LangSmithPer-release pass rate
02
Model registry + audit log
MLflow registry. Every inference logged with model version, input, output, and user.MLflow registryPer-inference logVersion pinningPlain-English trail
03
Idempotent tool calls
Idempotency-Key headers plus a dedupe ledger. Replay-safe. Double-charges architecturally impossible.Idempotency-KeyDedupe ledgerReplay-safe retriesNo double-charge
04
Drift + cost monitoring
Evidently or WhyLabs plus per-tenant cost attribution. Alerts before the customer or the CFO notice.Evidently / WhyLabsPer-tenant costDrift alertsSpend ceilings
05
Multi-provider failover
OpenAI, Anthropic, Bedrock, and Azure routing. No single-vendor failure mode on the critical path.OpenAI + AnthropicAWS BedrockAzure OpenAILiteLLM gateway
06
AI off-switch
The product earns its keep without the AI tier. We design the fallback before we wire the model.Graceful degradationDeterministic fallbackFeature flagsQuota guardrails
Twelve to sixteen weeks. Eval-gated from day one.
5-day audit
Customer wedge, model choice, regulatory shape. Eval plan and architecture brief drafted. Sprint quoted at audit end.
Eval harness first
Golden Q&A set, CI gate, model registry stood up before a single feature. The pass rate exists before the product does.
Core path + tools
The one job your first customer pays for, end to end. Idempotent tool calls and the off-switch wired as we go.
Failover + drift
Multi-provider routing, drift and cost monitors, per-tenant attribution. The 3am questions answered before 3am.
Hardening + load
Held-eval pass rate locked. Cost ceilings set. Load tested at the scale your Series A partner will ask about.
Launch, eval published
Real customers, real inference, real audit trail. Pass rate published. Not a demo, not a friends-and-family beta.
Handover + appendix
Documented codebase, decision log, and the term-sheet appendix your investor reads first.
The stack we ship AI startups on
Boring on the spine. Modern where it earns its keep.
We lead with the stack a Series A partner trusts on sight, layer the AI gateway and eval tooling on top, and reach into the infrastructure tier only where scale justifies it. Nothing experimental on the critical path.
Tier 1 · What we build on
React + Next.js, Node, MongoDB, Postgres.
TypeScript throughout. The default for AI-first SaaS and dashboards.
AI layer · Models + orchestration
OpenAI, Anthropic, LangGraph, LiteLLM.
Multi-provider gateway with Bedrock and Azure failover behind it.
Retrieval · Vector + search
pgvector, Pinecone, Weaviate.
Citation-grade retrieval. No hallucinated sources in production.
Eval + observability
Braintrust, LangSmith, MLflow, Evidently.
The pass rate, the registry, the drift monitor. Your audit trail.
When the brief asks for it
Python for inference, data, and pipelines.
Spark and Kafka where the data volume actually warrants it.
Infrastructure that scales it
AWS, Kubernetes, Docker, Redis, Terraform.
Comfortable on GCP, Azure, Vercel, Fly.io. We deploy where your team already is.
Scale + receipts · what your CFO and VC ask about
Numbers that survive diligence. Not adjectives.
When your acquirer’s CTO or your Series B partner walks the diligence, they read figures, not a deck. These are the receipts that hold up under questioning, drawn from seven years of production code we wrote and still maintain.
Production scale
Cumulative GMV and transactions through code we shipped, with API calls per year on production code we wrote. Zero customer-data breaches in production across seven years.
£540M+
GMV / transactions
320M+
API calls / yr
Funded founders
Series A founders we built with who went on to raise Series B, an 84% cohort rate. Two clients reached IPO with code we wrote. Three reached an M&A exit.
12
raised Series B
2 IPO
3 M&A exits
Trust + continuity
SOC 2 Type II achieved at Empyreal in December 2025, held pen test. Median client tenure on retainer of 28 months. The same seniors sign, ship, and maintain.
SOC 2
Type II, Dec 2025
96%
engineer retention
Start with the audit. The sprint is quoted at the audit’s end.
You don’t commit to a sixteen-week build on a first email. You start with a fixed-price 5-day audit, see the eval plan and architecture brief, then decide. Pricing in the reply, structured around your runway.
£8K fixed. The eval plan, the architecture brief, the costed sprint.
Two senior engineers read your wedge, model choice, and regulatory shape, then hand back a one-page architecture brief, an eval plan, and a costed sprint. Often signed standalone before the bigger build.
- Eval plan + held-eval design
- One-page architecture brief
- Costed, fixed-scope sprint quote
- 5-day AI startup audit · £8K fixed
The full AI startup build, fixed scope, eval-gated.
Twelve to sixteen weeks, two senior engineers paired daily. The six artefacts shipped into the product. Milestone payments structured around your runway and round timeline, not ours.
- Eval suite + model registry + audit log
- Idempotent tools + multi-provider failover
- Term-sheet appendix at handover
- Eval-gated sprint · from £65K
A second senior voice on the AI calls that matter.
For founders with an in-house team. We sit alongside on the eval strategy, the provider gateway, the cost engineering on inference. The call before the next architecture decision is made alone.
- Fortnightly architecture review
- Eval + cost-engineering on inference
- Clean exit, 30-day walk-away both ways
- Advisory or embedded senior · from £5K/mo
Procurement, security, and trust. What your CFO, CTO, and DPO read in writing.
When your first enterprise prospect sends the 60-page security questionnaire, you don’t panic. The answers already exist in the repo. Six artefacts your diligence call leans on, every engagement.
SOC 2 Type II achieved.
Empyreal-internal SOC 2 Type II achieved December 2025. A Vanta or Drata evidence pipeline wired on day one for your build.
Signed DPA + SCCs.
UK and EU GDPR-aware DPA. SCCs for cross-border. A sub-processor list, versioned and procurement-tested.
DSAR runbook + erasure.
A 30-day DSAR response that’s defendable. Right-to-erasure at the schema. ICO audit-defendable from launch.
30-day walk-away, both ways.
You keep everything. IP assigns on commit. Two handovers in seven years, both inside 48 hours.
No-poach for agencies.
White-label friendly. A 12-month no-poach with referrer agencies on request.
Procurement security pack.
SIG-Lite and CAIQ pre-filled. Pen-test report. Architecture diagram. DR plan. SLA and uptime SLO.
Three AI builds, three outcomes
Anonymised by contract. Verifiable on a call.
AI-first revenue ops, rebuild
Seed MVP with no eval, no audit log. We rebuilt in 14 weeks with 240 golden pairs gated in CI and Bedrock plus Azure failover.
Citation-grade doc intelligence
Founder needed a chatbot that never hallucinates a source. pgvector retrieval, eval gate at 94% held-eval pass, idempotent tool calls.
Clinical triage assistant, diligence-ready
UK GDPR plus NHS DSPT alignment. Per-inference audit log, AI off-switch, deterministic fallback. The IG team approved on first review.
What founders ask before they sign.
You start with a fixed £8K, 5-day AI startup audit. At the end you get a costed, fixed-scope sprint, typically from £65K for a 12-to-16-week build. No open-ended hourly meter, no surprise change orders. The audit fee comes off the sprint if you go ahead.
Yes. AI startup development here starts with the eval harness, not the features. We stand up a held-eval golden set and a CI regression gate in the first three weeks, then publish the pass rate per release. The number goes in your term-sheet appendix. Our median across AI builds shipped is 94% at launch.
Nothing they notice on the critical path. We route across OpenAI, Anthropic, Bedrock, and Azure, so there’s no single-vendor failure mode. And the AI off-switch means the product still earns its keep on a deterministic fallback if every provider is down at once.
Tell us early. Milestone payments are structured around your runway and round timeline, not ours. Several founders have raised a few weeks late in the last 18 months and all of their sprints finished. There’s a 30-day walk-away clause both ways, so you’re never trapped.
That’s the whole point of the build. You get a model registry with a per-inference audit log, idempotent tool calls, drift and cost monitoring, and an architecture brief. Twelve Series A founders we built with went on to raise Series B. Two clients reached IPO on code we wrote.
The answers ship in the repo. SIG-Lite and CAIQ come pre-filled, with a pen-test report, an architecture diagram, a DR plan, and a signed DPA with SCCs for cross-border. Empyreal achieved SOC 2 Type II in December 2025, and we wire a Vanta or Drata evidence pipeline on day one of your build.
Two senior engineers pair daily on every AI startup build, so the context lives in two heads and the decision log, not one. Engineer retention here is 96%, the same seniors sign, ship, and maintain. If anyone moves on, the documented codebase means a new engineer is productive in their first fortnight.
UK and EU GDPR-aware from the schema up. Signed DPA, SCCs for cross-border transfers, a versioned sub-processor list, a 30-day DSAR runbook, and right-to-erasure built into the data model. ICO audit-defendable from the day you launch.
In context
Inside the work.
A look at the kind of ai startup development surface we hand over — real screens, real data, documented and yours from day one.